Optimal Aggregation of Prediction Intervals under Unsupervised Domain Shift

As machine learning models are increasingly deployed in dynamic environments, it becomes paramount to assess and quantify uncertainties associated with distribution shifts. A distribution shift occurs when the underlying data-generating process changes, leading to a deviation in the model's performance. The prediction interval, which captures the range of likely outcomes for a given prediction, serves as a crucial tool for characterizing uncertainties induced by their underlying distribution. In this paper, we propose methodologies for aggregating prediction intervals to obtain one with minimal width and adequate coverage on the target domain under unsupervised domain shift, under which we have labeled samples from a related source domain and unlabeled covariates from the target domain. Our analysis encompasses scenarios where the source and the target domain are related via i) a bounded density ratio, and ii) a measure-preserving transformation. Our proposed methodologies are computationally efficient and easy to implement. Beyond illustrating the performance of our method through real-world datasets, we also delve into the theoretical details. This includes establishing rigorous theoretical guarantees, coupled with finite sample bounds, regarding the coverage and width of our prediction intervals. Our approach excels in practical applications and is underpinned by a solid theoretical framework, ensuring its reliability and effectiveness across diverse contexts.

Paper

References (49)

Scroll for more · 37 remaining

Similar papers

Peer review

Reviewer EZSx6/10 · confidence 2/52024-07-06

Summary

It is crucial to assess and quantify uncertainties associated with distribution shifts in the context of unsupervised domain shift. The paper propose methodologies for aggregating prediction intervals, which capture the range of likely outcomes for a given prediction. The authors support these methodologies theoretically by considering a bounded density ratio and a measure-preserving transformation between source and target domains. This includes finite sample bounds for the coverage and width of their prediction intervals.

Strengths

1.The method is innovative in aggregating various prediction intervals on the target domain under the domain-aligning assumption within an unsupervised domain adaptation framework. 2.The paper includes sufficient theoretical discussions, presenting finite sample concentration bounds to demonstrate that their method achieves adequate coverage with a small width.

Weaknesses

1.Apart from theoretical results, there is a lack of intuitive understanding as to why aggregation can achieve minimal-width prediction intervals. 2.The paper does not include comparisons with previous methods of aggregated prediction intervals (Hosen et al., 2014). 3.Limited experiments and comparative methods make it difficult to affirm the effectiveness of the algorithm. [1] Hosen, A., Khosravi, A., Nahavandi, S., and Creighton, D. Improving the Quality of Prediction Intervals Through Optimal Aggregation. IEEE Transactions on Industrial Electronics. 2014

Questions

1.What does $f \in \mathcal{F}$ signify in equation (2.1)? There is no $f$ explicitly used in the main part of this equation. 2.What does $m_0$ represent in line 118? Is it the same as $m_0$ mentioned in line 137? What is the relationship between $m_0$ and $f_0$? 3.In line 164, it states, "However, any such estimator $ \hat{\omega}(x) $ can be non-zero for $ x $ where $ \omega_0(x)=0 $ due to estimation error. Consequently, $ \hat{\omega} $ may not be efficient in selecting informative source samples." Given this observation, why not apply the hinge function not to $ \hat{\omega} $, but to $ (Y - m_0(x))^2 - f(x) $ in Equation (3.5)? 4.The notation $B_{\mathcal{F}}$ first appears in Theorem 3.2 (please correct me if I am mistaken), and there is no explanation provided for it. 5.The paper introduces two algorithms, but Section 5 includes experiments for only one of them. Which algorithm is it, and why are there no experiments conducted on the other one?

Rating

6

Confidence

2

Soundness

3

Presentation

3

Contribution

3

Limitations

N/A

Reviewer zRWP5/10 · confidence 3/52024-07-10

Summary

The authors study the problem of how to construct prediction intervals on a target domain under both covariate shift and domain shift assumptions (i.e., the source and target domains are related either via a bounded density ratio, or a measure-preserving transformation), designed to ensure adequate coverage while minimizing interval width. They provide theoretical guarantees, with finite sample bounds, regarding the prediction interval coverage and width. They apply their method on the airfoil dataset, comparing the performance with weighted split conformal prediction.

Strengths

- The paper is well-written overall, the authors definitely seem knowledgeable. - The paper studies an important and interesting problem. - The proposed method outperforms weighted split conformal prediction on one dataset.

Weaknesses

- The experimental evaluation is limited to a single low-dimensional tabular dataset, with a synthetic/simulated distribution shift. It is not possible to determine if the proposed method actually would have significant real-world utility/impact. Summary: - Important problem and a well-written paper overall, but the experimental evaluation is too limited. I think this could be a solid paper, but I don't think it can be accepted in its current form.

Questions

- Could you extend the experimental evaluation with more datasets and baseline methods? - Why is the figure on page 9 missing a figure number and caption?

Rating

5

Confidence

3

Soundness

2

Presentation

3

Contribution

2

Limitations

Yes.

Reviewer zRWP2024-08-13

_(I will quickly write a reply now, to try to give you some time to respond before the deadline, but I will also think more carefully about this later before making my final recommendation. I will still consider raising my score)_ *** *** > (notably, the AssetWealth dataset you referenced is 13GB) - There are 8 different datasets in the TMLR paper i pointed to, I thought that maybe it could be easy to apply to at least one of those? > ...suggesting its significant potential impact across diverse real-world scenarios - The method has only been applied to a single dataset with a real-world distribution shift though, right? Are there some other tabular datasets with realistic distribution shifts? > These experiments consistently demonstrate the advantages of our method over WVAC and WQC - Expect for "Airfoil data with Algorithm 2" where the coverage drops below 0.95? Also, is this the only evaluation of Algorithm 2? Why? (could perhaps be made more clear in general what the two different algorithms should be utilized for, how do I know which one to use in a real-world application?) > ...performing dimensionality reduction. Once these preprocessing steps are completed and we obtain vector representations of the images, our method can be implemented efficiently and executed rapidly - Why is dimensionality reduction a required step? Isn't this a quite significant limitation of your method then, which would limit the real-world utility also for non-image data? No dataset in the experiments has more than 8(?) features, what's the maximum number your method can handle?

Authorsrebuttal2024-08-13

Thanks for your further comments. 1. **The method has only been applied to a single dataset with a real-world distribution shift though, right? Are there some other tabular datasets with realistic distribution shifts?** Following your concern, we managed to run an additional experiment within this short timeframe. The new dataset is the *Appliances Energy Prediction Dataset*, which is freely available from the UCI repository. This dataset is a time series data with 28 covariates and one response variable. We used data from 2016-01-11 to 2016-02-15 (5000 samples) as our training set and data from 2016-05-13 to 2016-05-27 (2000 samples) as our testing set. Since the source and target data are from different time periods, this experiment involves a non-synthetic real-world time shift. The results based on our Algorithm 2 are presented below: **Table:** Experimental results for the Appliances Energy Prediction Dataset | Outcome | Our Method | WVAC | WQC | |-----------|------------|----------|---------| | Coverage | 0.95 | 1.00 | 1.00 | | Bandwidth | 461.69 | 6809.87 | 2032.12 | 2. **Why is dimensionality reduction a required step?** We would like to clarify that our experiment involves two main steps: first, training the component functions, and second, efficiently aggregating these components. Our methodology specifically focuses on the aggregation step, where we combine pre-trained models to obtain a prediction interval with adequate coverage and minimal width. The key point is that our method assumes these component functions are already trained, and our contribution lies in the efficient aggregation of these components. When we say our method is easy to implement, we are referring to this aggregation step, which is independent of the data's dimensionality. In many real-world applications, these components are typically pre-trained and provided to us. However, we acknowledge that if the pre-trained components are not available, they must first be trained from the data. This training process can indeed be time-consuming, especially as the models' complexity increases. The dimension reduction step may be required for training these component functions, but is **not required** for aggregation. To summarize, when we refer to *our method*, we mean the aggregation (shape estimation) and shrinkage steps, which are straightforward to implement and do not depend on the data’s dimensionality. The more challenging aspect is the initial training of the components if they are not already available. Our method can indeed handle large datasets, with the main consideration being the time required to compute the component functions. To illustrate, we conducted an additional experiment on Appliances Energy Prediction Datase, using data with 28 covariates. It's important to note that the aggregation step involves combining multiple component functions, each of which is univariate. Therefore, the dimension involved in the aggregation step is the number of component functions, **not the dimension of the data**; the data's dimensionality only affects the training of the component functions. 3. **Expect for "Airfoil data with Algorithm 2" where the coverage drops below 0.95? Also, is this the only evaluation of Algorithm 2? Why?** As we mentioned, the drop in coverage below 0.95 is due to the small sample effect rather than a limitation of our methodology. We evaluated Algorithm 2 on both the Airfoil data and the ETDataset. The choice between Algorithm 1 and Algorithm 2 depends on prior knowledge about the type of distribution shift. If the shift is due to covariate changes, and the support of the test covariates lies within the support of the source covariates, then Algorithm 1 is appropriate. However, Algorithm 2 is more suitable when the shift is caused by transformations (e.g., shifts, rotations). If the support of the source and target covariates is likely to be (almost) disjoint, density ratio estimation will be poor, making Algorithm 2 the better option in such cases. 4. **There are 8 different datasets in the TMLR paper I pointed to, I thought that maybe it could be easy to apply to at least one of those?** All eight datasets referenced in the TMLR paper are large-scale datasets. Even the smallest among them, SkinLesionPixels, is still 2GB. Additionally, accessing some of these datasets requires account registration, which further complicates the process.

Reviewer zRWP2024-08-14

Good clarifications, thank you. And thanks for the detailed responses throughout, I appreciate the effort put into this. I still really would have liked to see the proposed method being applied to non-tabular datasets, but I'm now leaning towards accept either way. I will raise my score to "5: Borderline accept".

Authorsrebuttal2024-08-14

Official comment by Authors

Thank you very much for your suggestions and responses throughout the review session! We believe this has truly helped our paper, and we will add all the additional details to the revised version of the manuscript.

Reviewer 88pD6/10 · confidence 4/52024-07-10

Summary

Building on the work of Fan et al. (2023), this paper addresses the challenge of computing prediction intervals in an unsupervised domain adaptation setting, where labeled samples are available from a related source domain, and unlabeled covariates are available for the target domain. The primary objective is to generate prediction intervals for the target domain that provide both an asymptotic coverage guarantee and minimal width. The authors also explore the aggregation of prediction intervals on the target domain under both covariate shift and domain shift scenarios. Theoretical guarantees, including finite sample concentration bounds on approximation errors and coverage guarantees, are presented. The proposed methods are illustrated and compared with the weighted split conformal prediction method using a real-world dataset.

Strengths

- The paper tackles the important and challenging problem of uncertainty quantification in the unsupervised domain adaptation setting. - The authors propose algorithms to generate prediction intervals with both coverage and minimal width guarantees. - Experimental results show that the proposed method generates prediction intervals with comparable coverage but smaller average width compared to an existing method.

Weaknesses

- The paper assumes familiarity with Fan et al. (2023), making it difficult for readers who have not read that work to understand some notations and concepts. Section 2.1 on problem formulation is particularly hard to read without prior knowledge. - There is a lack of flow in the paper, with some explanations and definitions missing or unclear. For instance, terms like "valid prediction interval", and expressions such as U(x) and l(x) are not clearly explained. - The authors write "This translates into solving the following optimization problem:" without any explanations. - Another example is Line 234, "adding a slight delta to ensure coverage even when F is complex.". This is not clear. It is discussed in Fan et al (2023) in (2.4). - Even if simple, I suggest adding important details about different expressions in the appendix to help the reader. - The title mentions "optimal aggregation," but model aggregation is not the central focus of the paper. It is first discussed in Remark 4.2 and briefly mentioned in line 272. - The paper should highlight the similarities and differences with Fan et al. (2023) in terms of contributions, both theoretical and empirical. Also, the authors use different quantile levels for estimators compared to Fan et al. (2023) without explaining the reason for these differences. In Fan et al (2023), the following quantile levels are considered: 0.05, 0.35, 0.65, and 0.95, while the authors use 0.85, 0.95, 0.9, and 0.9. - The statement that "the shape of the prediction band does not change much if we change the level of coverage" lacks clarity on scenarios where this might not be true, such as multimodal distributions (?). - The discussion on lambda <=1 is confusing and needs improvement. The paper inconsistently explains the shrinkage factor and its implications. For example, the authors write "Therefore, it is not immediate whether the shrinkage factor is smaller than 1, i.e., whether we are indeed shrinking the confidence interval (lambda > 1 is undesirable, as it will widen finit , increasing the width of the prediction band)." and "Although ideally lambda <= 1, it is not immediately guaranteed as we use separate data for shrinking.". The authors also discuss it in Lemma 3.3. and Lemma 4.4. - The term "shape of the prediction band" with prediction intervals can be confusing as it can have other meanings (see functional data). I suggest the authors clarify this in line 147. - The paper ignores a significant amount of research on prediction interval estimation. I suggest discussing important related work. - The limitations of the paper are not discussed. For example, focusing solely on intervals might have implications under multimodal distributions, where High-Density Regions (HDR) would be the smallest regions. - The relationship between the proposed approach and conformal prediction is not clearly explained. The paper should discuss the guarantees provided by the proposed approach compared to conformal prediction, including the relevant theoretical guarantees and calibration aspects. For example, the weighted conformal approach also splits the data and estimates a density ratio. What about the (weaker?) guarantees you provide? The following paper could also be useful: "On the expected size of conformal prediction sets". Also, in conformal prediction, in addition to finite-sample marginal guarantees, the authors often provide asymptotic conditional guarantees. What can you say about this with your approach? - The comparison with the conformal approach in the experiments might not be fair, as conformal prediction does not optimize for the smallest width. Additionally, reporting the interval score, which is a proper scoring rule for intervals, would be useful for evaluating the methods. - Typos and clarifications: - line 137: "remains same." - line 154: "via by solving" - line 453 "a.s. on source". On the source domain? - line 470, "we with have" - It is not "Leduox-Talagrand" but Ledoux-Talagrand". - line 481, a minus sign is missing - Line 490 to 492. It is not clear why tilde tau <= alpha. - Line 182, what does \mathcal{F} - f* mean? - About Theorem 3.4, we expect the performance (or convergence rates) to also depend on alpha. Where does this appear in the theoretical guarantees?

Questions

See weaknesses.

Rating

6

Confidence

4

Soundness

3

Presentation

2

Contribution

3

Limitations

The limitations are not discussed in the paper.

Reviewer 6ToN5/10 · confidence 4/52024-07-12

Summary

This work proposes a method to construct prediction intervals in distribution shift problems, where unlabeled data from the target domain is available. The method is inspired by Fan et al (2023) in the i.i.d. setting and assumes we can estimate a transform (defined by reweighting or a transport map) that maps the joint distribution from the source domain to the target.

Strengths

- The problem studied is relevant. - The method appears reasonable and sound.

Weaknesses

- My main concern is about the novelty of this work: given the assumed condition on domain shift (namely a good approximation to the reweighting or transport map is readily available) it is a natural application of the method in Fan et al (2023). The proof will require some additional effort but given the assumptions I do not expect them to be technically challenging. - While the paper employed assumptions on domain shift are common in literature, I am uncertain about their relevance, especially in high-dimensional problems. It is well-known that the bounded density ratio assumption in Section 3 is unrealistic in high dimensions. And while I have come across past works on the transport map assumption, I do not know of realistic examples where the assumption actually holds so that the coverage guarantees established in this work could be relevant: it is much easier to think of scenarios where the assumption does not necessarily hold, but heuristic methods based on this assumption achieve better predictive performance; the latter is more often the goal in previous works. - Finally, I have doubts about the novelty/utility of Fan et al (2023), which seems quite similar to conformal prediction. It could be alternatively summarized as applying quantile estimation to the "conformity score" of $(y - \hat m(x)) / f(x)$, and compared with conformal prediction the main difference is the lack of sample splitting (in split conformal) or leave-one-out operations (as in full conformal). I wonder if the removal of sample splitting is really a wise choice: the coverage guarantees in Fan et al (2023) and this work relies on the Rademacher complexity of the function class for $f$. If the function class is defined by general ML models the guarantees would be rather unreliable in high-dimensional problems, in sharp contrast to the distribution-free guarantees in (split/full) conformal prediction. Otherwise, we would need further sample splitting to construct a finite set of candidates for $f$, in which case we could just do split conformal prediction.

Questions

Given the above questions on the theoretical motivation, novelty or significance, I believe there should be extensive empirical evaluation for the proposed method. In light of question 3 above, I also wonder if the observed performance difference between the proposed method and conformal prediction could be attributable to the choice of different conformity scores; how does the method compare to an implementation of split conformal prediction using the same $(y - \hat m(x)) / f(x)$ as the conformity socre?

Rating

5

Confidence

4

Soundness

3

Presentation

3

Contribution

2

Limitations

Yes.

Authorsrebuttal2024-08-12

As the rebuttal period is coming to a close, we wanted to check if there are any remaining concerns that might be preventing you from adjusting your score. If there is any additional clarification we can provide, please let us know. We would be more than happy to address any questions you might have. Thank you again for your time and effort in reviewing our paper.

Reviewer zRWP2024-08-08

Thank you for your response. I have read the other reviews and all rebuttals. The other reviews are all borderline, but slightly more positive than mine. The authors add some new experiments, and respond well overall to the reviews. My main concern _"The experimental evaluation is limited to a single low-dimensional tabular dataset, with a synthetic/simulated distribution shift. It is not possible to determine if the proposed method actually would have significant real-world utility/impact"_ is not fully addressed. The evaluation is still limited to low-dimensional tabular datasets with synthetic/simulated distribution shifts, and I thus still find it difficult to know if the method actually would have significant real-world utility. - Could you evaluate on some dataset with a real-world distribution/domain shift? - Could you evaluate on some non-tabular dataset, on an image-based dataset? - Also, is it not a fairly major concern that the coverage of your method falls below 0.95 in the "Airfoil data with Algorithm 2" experiment? Yes, the interval length is roughly half that of WVAC and WQC, but is this (big) improvement worth it if the method fails to achieve valid coverage? I am open to increasing my score at least to "5: Borderline accept", but would still like to see at least some more convincing experimental results.

Authorsrebuttal2024-08-10

Thanks for your further comments. 1.**Could you evaluate on some dataset with a real-world distribution/domain shift?**: We applied our method to the ETDataset (ETT-small) from [1], which contains hourly-level data from two electricity transformers at two different stations, including measurements of load and oil temperature. Each data point consists of 8 features, including the date of the point, the predictive value "oil temperature", and 6 different types of external power load features. For our experiment, we used the data from one transformer during the period from July 1, 2016, to November 2, 2016, as our source data, and data from the same time period from the other transformer as our target data. As our source data and the target data are from different locations, we have a geographical covariate shift; see [1] for more details. Our results are as follows: **Table:** Experimental results for the ETDataset | Outcome | Our Method | WVAC | WQC | |-----------|------------|-------|-------| | Coverage | 0.976 | 0.982 | 0.842 | | Bandwidth | 41.525 | 57.9 | 54.981| [1] Haoyi Zhou, etc. Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting. 2. **Could you evaluate on some non-tabular dataset, on an image-based dataset?** The widely used non-tabular data in the field of distribution shift are various types of image data (e.g., MNIST-USPS digit data, Waterbird data etc.), which are generally used for classification tasks. However, our current method for constructing prediction intervals relies on the continuous response variable. Therefore, it cannot accommodate non-continuous responses. We acknowledge this limitation of our method and believe that this extension will be an interesting future research direction. 3. **Also, is it not a fairly major concern that the coverage of your method falls below 0.95 in the "Airfoil data with Algorithm 2" experiment? Yes, the interval length is roughly half that of WVAC and WQC, but is this (big) improvement worth it if the method fails to achieve valid coverage?** We believe the reason our method fails to achieve coverage is a small sample effect rather than the limitation of our methodology. As is evident from other real data experiments, our method achieves the nominal level of coverage. Therefore, our method *does not sacrifice coverage to reduce width*; our theoretical results indicate that asymptotically, our method achieves adequate coverage while having minimal width.

Reviewer zRWP2024-08-11

Thank you. For the image-based data, would it perhaps be possible to apply your method to one of the regression datasets from _How Reliable is Your Regression Model's Uncertainty Under Real-World Distribution Shifts?_ (TMLR, 2023)?

Authorsrebuttal2024-08-12

Thank you very much for bringing this benchmark paper to our attention for evaluating uncertainty estimation methods. Since your last comment, we have made a concerted effort to apply our method to the data referenced in the paper. However, due to limited computational resources and the large scale of the data (both in terms of high dimensionality and the number of training and testing samples), we found it challenging to process the datasets within the short timeline available. Specifically, we encountered difficulties running the ``create\_datasets.py" script, embedding the images into vector space, and applying dimension reduction techniques to extract useful information from the images. Nevertheless, we will cite the paper you mentioned as a valuable source of non-tabular data with distribution shifts, which could be used to evaluate our methods in future work. We would also like to take this opportunity to highlight the key contributions of our paper: 1. **Methodology:** We propose two novel methods (Algorithm 1 and Algorithm 2) to address distribution shift challenges where no labels from the target domain are observed. Our approach is straightforward to implement and involves solving a convex optimization problem. 2. **Theoretical Guarantees:** We establish rigorous theoretical guarantees, including finite-sample concentration bounds, to demonstrate that our method achieves adequate coverage with a small width. 3. **Experiments:** We evaluate Algorithm 1 on three real-world datasets: the Airfoil data, Real Estate Dataset, and Energy Efficiency data. Additionally, we assess Algorithm 2 on the Airfoil data. We further demonstrate the utility of our method by applying it to the ETDataset, which presents a non-synthetic geographical covariate shift. Finally, we highlight the advantages of our method compared to conformal prediction through robustness checks. We sincerely hope that our paper's contributions are properly recognized and that our responses have addressed most of your concerns. Thank you again for your valuable feedback. Please do not hesitate to let us know if there are any other concerns that might be preventing you from adjusting your score.

Reviewer zRWP2024-08-13

Thank you again for the response. While I realize that you don't have a lot of time to implement new experiments, I do still find it quite problematic that it hasn't been demonstrated that the proposed method actually can be applied to image-based datasets. My main concern _**"It is not possible to determine if the proposed method actually would have significant real-world utility/impact"**_ therefore mostly remains. Also, this doesn't quite seem to align with your statement _"Our approach is straightforward to implement"_ from above. You have however provided a detailed rebuttal, I can appreciate that there are some other contributions of the paper, and the other reviewers are now all positive overall. I have raised my score to "4: Borderline reject" for now.

Authorsrebuttal2024-08-13

Thank you for your feedback and for the increased score. We appreciate your concerns and would like to address them in the following: 1. **Real-World Utility**: While we were unable to demonstrate the utility of our method on image-based data within the limited time available (notably, the AssetWealth dataset you referenced is 13GB), it's important to highlight that numerous real-world applications extend beyond image-based tasks. Examples include price prediction in financial markets, weather forecasting, and energy consumption analysis, among others. Given the vast array of real-world tasks, it's impractical to test our method on all possible datasets. However, we've showcased its utility through experiments on the Airfoil dataset (sound pressure prediction), Real Estate Dataset (house price prediction), Energy Efficiency data (heating load prediction), and ETDataset (oil temperature prediction). These experiments consistently demonstrate the advantages of our method over WVAC and WQC, suggesting its significant potential impact across diverse real-world scenarios. Additionally, ***the primary contributions*** of our paper lie in its methodological and theoretical advancements, particularly the finite sample coverage and width guarantees, as previously mentioned. 2. **Implementation Ease**: Our approach is straightforward to implement in that it primarily involves solving convex or linear programming problems. However, processing large datasets, especially image-based ones, can be inherently time-consuming due to steps like embedding images into vector spaces and performing dimensionality reduction. Once these preprocessing steps are completed and we obtain vector representations of the images, our method can be implemented efficiently and executed rapidly. We genuinely value your insights and believe that our method holds promise for various applications. We would greatly appreciate it if you could consider further raising your score above the acceptance threshold if you find our clarifications satisfactory.

Reviewer EZSx2024-08-10

Thank you for your detailed explanations. I appreciate the way you’ve clarified the intuitive understanding and your commitment to comparing it with previous approaches. However, in the quesiton 3, it states that "$\hat{\omega}$ may not be efficient in selecting informative source samples." Given this, would applying the hinge function to $(Y - m_0(x))^2 - f(x)$ help address this issue? Besides theoretical analysis convenience, are there any other benefits to this approach?

Authorsrebuttal2024-08-10

Thanks for your further comments. In Step 1 (Shape estimation), the ideal optimization problem we aim to solve is given by equation (3.2) in our paper, which can be expressed as: $$ \\min\_{f \in \\mathcal{F}} \ \mathbb{E}\_{n, T}[f(X)]\, \ \ {\rm s.t.}\ \\mathbb{E}\_{n,S}[w_0(X)\\mathbf{1}_{\{f(X)<(Y-m_0(X))^2\}}]=0 \.\tag{1} $$ If we had perfect knowledge of the source samples for which $ w_0 > 0 $, we could directly solve equation (3.2). However, in practice, $w_0$ is unknown, and we must estimate it using $ \hat{w} $. This introduces the possibility of estimation error, where $ w_0(X_i)=0 $ but $ \hat{w}(X_i)> 0 $ for some source samples $X_i $. Such errors lead to additional constraints—namely, requiring $ f(X_i)\geq(Y_i-m_0(X_i))^2 $ for all $X_i $ where $ \hat{w}(X_i)> 0 $—that are not necessary if $w_0(X_i)=0 $. To address this, instead of enforcing the right-hand side of the constraint in (1) to be exactly 0, we relax it to $ \epsilon $ to accommodate the estimation error in $\hat{w} $. This adjustment leads to the following optimization problem: $$ \min_{f \in \\mathcal{F}} \ \ \\mathbb{E}_{n, T}[f(X)]\, \ \ \text{s.t. }\ \\mathbb{E}\_{n,S}\left[\hat{w}(X)\\mathbf{1}\_{\{f(X)<(Y-m_0(X))^2\}}\right]\leq \epsilon \.\tag{2} $$ However, there is a caveat: the presence of the indicator function in (2) makes the constraint non-convex. To address this, we replace the indicator function with a convex surrogate, specifically the hinge function $ h_{\delta}\left((Y-m_0(X))^2-f(X)\right) $, which restores convexity to the problem. In summary, we start with equation (3.2) from our paper, substitute $w_0 $ with $ \hat{w} $, and then replace the indicator function with a hinge function to ensure the problem remains convex. Thus, the use of the hinge function is driven by computational considerations rather than merely for theoretical analysis.

Reviewer EZSx2024-08-12

Thanks for your detailed repsonse. I have decided to increase the rating.

Authorsrebuttal2024-08-12

Thank you very much for taking the time to read our rebuttal. We greatly appreciate your positive feedback and the decision to increase the score.

Authorsrebuttal2024-08-10

Thank you very much for both your valuable reviews and the increased score. We will carefully revise our manuscript based on your suggestions.

Reviewer 6ToN2024-08-13

Thank you for your response. I appreciate the clarifications regarding the theoretical contributions and the new experiments. I will update my score accordingly.

Authorsrebuttal2024-08-13

Thank you very much for reading our rebuttal and for increasing the score. We will put the new experiments in our revision.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC